Hong Kong Exchanges and Clearing (HKEX) aims to transition to T+1 settlement in the second half of 2028.

Jupiter official announced that its high-leverage trading product YOLO has surpassed $1 billion in trading volume less than 24 hours after launch. The product currently supports up to 1000x leverage for SOL, BTC, and ETH. Developed by Jupiter Labs, YOLO is benchmarked against PaperTrade and adopts the JLP model. It simulates positions based on real market prices, with real profit and loss outcomes, but does not require executing corresponding large-scale leverage positions in the actual market. Profits are disbursed from the fund pool; in cases of insufficient liquidity, payouts may enter a queue.

Ronin officials announced that the migration of Ronin Waypoint has entered its final "grace period", running until 23:00 Beijing Time on October 31, 2026. Once the grace period concludes, Ronin Waypoint will be permanently shut down, marking the completion of the migration. Earlier, Ronin required users to transfer assets from their Waypoint wallets to Ronin Stash or mnemonic wallets by October 16, warning that untransferred assets may become inaccessible.

Beating AI Express News: Tencent Hunyuan announced the upgrade of its agent memory framework Hy Memory to version 2.0, and open-sourced the dedicated model Hy-Memory-30A3B, the memory framework itself, and the MemHub application. These three capabilities can be used independently or combined to form a complete memory solution. The framework handles the storage, update, and evolution of three types of memory—chat, work, and code—ranking first among open-source memory frameworks on four benchmarks. The dedicated model scored 55.49 in evaluations across multiple memory systems including Hy Memory, Mem0, and MemOS, surpassing Opus5’s 54.62. Hy Memory 2.0 upgrades chat memory from fragmented facts to complete event contexts, adds new work and code memory capabilities, and can judge the applicable conditions of experience to accurately assist in long-term tasks. As a local memory management app, MemHub supports CLI and Mac, connects to 10 AI tools including WorkBuddy, Cursor, and Codex, and enables historical memory import and real-time cross-tool synchronization. This capability has been deployed in Tencent’s business lines including Stock Selector, iWiki, and Tencent News: Stock Selector’s multi-hop recall rate rose from 50% to 90%, with a comprehensive score improvement of 23.8%; iWiki’s answer accuracy when using only memory increased by 48.6% relative to providing all original conversations; Tencent News achieved an average long-term preference score of 91.68 points.

OKX’s global managing partner and chief marketing officer Haider Rafique announced in a post that he will step down from his role and take a temporary break to focus on family and personal matters after nearly seven years at the crypto exchange. Insiders revealed Rafique plans to depart by the end of 2026. He joined OKX in March 2020, led the platform’s brand upgrade, and was involved in facilitating Intercontinental Exchange’s (ICE) $200 million strategic investment in OKX earlier this March.

According to Lookonchain monitoring, an address generated $10.1 billion in trading volume on the launch day of perpetual contract trading platform Papertrade, posting a $979,000 loss and receiving 252 million PAPER tokens. The address subsequently halted trading, staked the 252 million PAPER tokens to earn $1.73 million worth of USDC rewards, and has now netted a $762,000 profit.

Beating AI Express (Insight) – U.S. AI firm PrismML has rolled out three new models for its Bonsai 2 family: 2B, 4B, and 9B, all based on Alibaba’s Qwen3.5. The team restricted the language model’s core weights to only three values: -1, 0, and +1, cutting the model sizes to 0.43GB, 0.95GB, and 2GB respectively – roughly one-ninth the size of original 16-bit weight models. The new models retain support for inference, image understanding, and tool calling, and are primarily designed for local agents on devices like smartphones and personal computers. PrismML states that on the iPhone 18 Pro Max, the 9B model generates up to 34 tokens per second, while the 2B model hits 120 tokens per second. Official tests show the 9B model’s programming and instruction-following capabilities are nearly on par with the same-scale original Qwen3.5, though it lags significantly on complex tasks. In Tau2Bench, which assesses multi-turn dialogue and tool operations, Bonsai 2 9B scores 73.02, compared to the original model’s 85.25; its complex document parsing score is 73.8, versus 88.73 for the original. All three models have released their weights under the Apache 2.0 license.